KarthikKB1998/CARD-Audio-Captioning

CARD audio captioning demo

Interactive demo of "CARD: Cross-component Audio Representation Distillation for Encoder-Free Audio Captioning" (IEEE SLT 2026).

Pick a sample clip, listen to it, and reveal the caption written by `KarthikKB1998/CARD-Qwen3-4B-AudioCaps` next to the human reference captions from the test set.

This is a static page. The captions were generated beforehand by the released model with the paper's decoding settings (beam search, 4 beams, at most 40 new tokens). Beam search is deterministic, so they are the captions the model produces live for these clips. To caption your own audio, run the model in the Colab notebook.

CARD has no audio encoder at inference. A 13.2 M-parameter projector turns the log-Mel spectrogram into audio tokens, which a Qwen3-4B language model with merged LoRA adapters reads to write the caption. During training a frozen CLAP-HTSAT teacher was distilled into the model by component, its early stages into the projector and its later stages into the language model.

• —Code: https://github.com/KarthikKolluriKB/CARD
• —Each sample clip keeps its original source and license, shown with its caption.

@inproceedings{kolluri2026card,
  title     = {{CARD}: Cross-component Audio Representation Distillation for Encoder-Free Audio Captioning},
  author    = {Kolluri, Ganesh Pavan Kartikeya Bharadwaj and Zhang, Yuchen and Kampouridis, Michael and Shekhar, Ravi},
  booktitle = {Proceedings of the IEEE Spoken Language Technology Workshop (SLT)},
  year      = {2026}
}